equinox

he/him · FRRouting · Junisco & Hp3Com slayer · knows more about IP routing than is healthy · moonlights in bullying electrons around a circuit board

Public Key

npub12znsfznenq5pgwjulqpqeg9szhmq5kcuw40w2rzcv9vj4ugnuwaqsr4q3p

NIP-05 Address

equinox@chaos-social.mostr.pub

Profile Code

nprofile1qqs9pfcy3fues2q58fw0sqsv5zcptas2tvw82hh9p3vxzkf27yf78wspz4mhxue69uhhyetvv9ujuerfw36x7tnsw43q779jk2

Publishing to

relay.ditto.pub

Author Public Key

npub12znsfznenq5pgwjulqpqeg9szhmq5kcuw40w2rzcv9vj4ugnuwaqsr4q3p

Show more details

Last Notes

2026-10-30 09:11:48 UTC

On a completely random note, the shit the US pulled on Autistici/Inventati is absolutely rippling through the networking community, and people are actively discussing what to do about it… and against it.
The US will probably be able to pull these stunts a few more times, but hopefully not too many. There's work to be done.
https://assets.chaos.social/media_attachments/files/117/529/081/686/371/473/original/5eae911712146a58.jpg

2026-10-26 06:29:49 UTC

Congratulations (and Thanks!) to the Brazilian and Serbian voters for providing some of the most encouraging news we've had in a long time!

2026-10-20 21:55:00 UTC

With apologies to Randall Munroe [https://xkcd.com/221/] 😂
https://assets.chaos.social/media_attachments/files/117/475/452/254/443/495/original/9c1cfa2a8decaeb7.png

2026-10-08 20:10:25 UTC

cool… apparently the only QEMU stream channel (tty/serial/...) that doesn't lose data with FreeBSD guests when buffers are full is… … … …the usb-serial emulated FT232BM. (which is also quite slow.)
virtio console/serial port? hahahaha. send 4 MB, get ≈30 kB. Thanks 😑.
(This works perfectly fine with Linux guests. And I've tried every option & knob I could find.)

2026-09-22 21:05:46 UTC

@nprofile…mv7s feels like Tektronix saw what you're doing with the GPU in ngscopeclient… DPO71*AX got a nVidia T1000 GPU
[via https://www.youtube.com/watch?v=HV140Y8VU1s]
(of course it just makes sense, but I prefer my theory of them seeing your offload work 😂)

2026-07-28 20:48:18 UTC

God this AI shite I've been trying is just straight up braindead. And I don't think I'm willing to try and learn to use it "properly".
(I guess the GPU I'm running this on isn't good enough either [ROCm on AMD 6950 KXTX, only 16GB], and/or my model choice is shit [some hybrid composed thing I got pointed at, says gemma4 fable5 12B])
Not actually wanting to do any of this certainly doesn't help either. Maybe I should sink the time into the afl fuzzer instead…

2026-07-22 15:41:39 UTC

@nprofile…mv7s so @nprofile…xkrr is trying to debug weird behavior with a device's SFP+ ports that is suspected to be misconfig between operating (at 1G) as SGMII or 1000base-X (or some bits in the negotiation pages on either) - do you have any suggestions/ideas how to (cheaply) debug that?
Identifying a signal as SGMII vs. 1000base-X would be good, getting the AN pages dumped better…

2026-07-17 21:26:16 UTC

- reply

@nprofile…qww0 10-Minuten-slots? o.O

2026-07-02 12:25:50 UTC

- reply

@nprofile…mv7s the list of speed-ganged ports (and splitting options/limits) tends to be the first thing I (and many network engineers) look for in the manual. The second thing is the per slot power limit table & if there are high-power slots.

2026-07-02 12:22:33 UTC

- reply

@nprofile…mv7s that limitation on breakout speeds is _extremely_ common; it's generally a serdes quad with very limited clocking options. I'm a little surprised at this being flagged by you of all people 😂
On some devices it's even worse, AFAIR there's parts with 16 or 24 lanes sharing clock configs.

2026-06-15 07:24:13 UTC

- reply

@nprofile…mv7s my intuition says the second one, but I'm clueless…
I have, however, recently read an application note (don't remember context) that advises against serpentines like those on the first example due to "sideways" propagation. It suggested doing one long fold "back" instead. That does break the pair quite a bit more but they seemed to consider that less important. Maybe another option to simulate?

2026-06-09 13:17:28 UTC

- reply

@nprofile…3n63 that's another submission on the review pile 😂

2026-06-09 12:45:29 UTC

We have somehow ended up in a timeline where I have to review a conference submission about¹ Quantum Computing.
I have no clue about Quantum Computing.
🤨 😵‍💫
*Uses "request external review" form*
¹ technically, it`s about _interfacing to_ quantum computing, so it's not exactly QC itself. Still…

2026-05-14 15:38:35 UTC

Shopping list:
- smoke detectors
- fire extinguisher
(…don't ask…)
(…kitchen…)
I'm also appalled that Switzerland doesn't legally require smoke detectors in living spaces. I was aware I had none in this flat and that was already on the shopping list anyway.

2026-04-30 01:45:32 UTC

I don't understand why everyone is picking on algif_aead for copy.fail. It's authencesn that's actually fucking things up.
Sure, you could keep using authencesn for IPsec if you break AF_ALG AEADs. But why? I'd rather break ESN support in IPsec and keep AF_ALG working for AEADs.
It doesn't even break IPsec, it just makes it rotate keys more often in very high bandwidth situations.

2026-04-26 11:55:50 UTC

- reply

@nprofile…mv7s oh I'm not complaining about their charging for it. I'm complaining of the general weirdness of the system as a whole 😅

2026-04-26 11:48:43 UTC

- reply

@nprofile…mv7s RHEL is certainly "enterprise", for better and worse…

2026-04-26 11:47:29 UTC

- reply

@nprofile…mv7s ok, to be fair we never tried supporting Arch 😂
Our main problems with Fedora are actually weird packaging decisions for libraries we depend on, when they go through SOVERSION bumps. (That and the license server bullshit for RHEL. I have no idea how far that applies for different variants.)
But if you have users, it might be worth it. I guess for something like ngscopeclient, Fedora/RH might not be that rare to find as a desktop.

2026-04-26 11:42:42 UTC

- reply

@nprofile…mv7s do you have users on Fedora? In our experience it's the absolute worst in terms of time & resources needed to keep running vs. actual payoff in users. (Because both sides are notably worse than other distros.)
(We're considering switching to something else to cover rpm based distros, maybe Rocky Linux, but not sure yet)

2026-04-21 15:50:48 UTC

- reply

@nprofile…mv7s is that the Sonnet dongle or a more recent one? I thought about grabbing one, but tbh it's not worth it at this point.
As for putting things on TBT… there was some discussion about building a TBT OCP 3.0 NIC enclosure with some $people, that would also be a very nice form factor (not quite as small, but way smaller than PCIe card enclosures.)
Primary issue with that is OCP 3.0 expects good forced airflow…

2026-04-21 15:47:29 UTC

- reply

@nprofile…mv7s the alternative for the laptop would be a chonky PCIe enclosure, which I do have but keep for other tinkering. For docks with builtin 10GbE I unfortunately haven't found any with an SFP+ slot 😅
But also, yeah, 10Gbase-T is utter garbage. The SN2010 actually won't enable more than 4 Cu-SFP+s (in ports 1, 2, 17 and 18) because only those have 2.5W power budget. The other SFP28 ports have only 1.5W, and, well… 🫠

2026-04-21 15:19:27 UTC

- reply

@nprofile…mv7s I just noticed I can see on my power graph whether the 10GbaseT to my laptop is up 😂
(This meter only has the switch right now though, a Mellanox SN2010, I'm moving & stuff isn't set up yet)
https://assets.chaos.social/media_attachments/files/116/443/367/254/234/577/original/79f62ffed0b1367f.png

2026-04-21 09:26:14 UTC

- reply

@nprofile…mv7s 100W continuous for a month is $28 around here, it sums up 😂
But yeah if it's not outlet metered 🤷

2026-04-21 09:18:07 UTC

- reply

@nprofile…mv7s out of curiosity, do you have that switch on a metered circuit? If yes: how much does it pull idle? (And what's the exact type?)

2026-04-05 20:13:02 UTC

- reply

@nprofile…y6rq I don't think you're making the right argument there. There's simply no consensus for this document. There's a lot of cases where that's a hard determination to make; a pretty even split is not one of them. If anything, showing that it is in fact even (and not biased by random drive-bys or sockpuppets), and then insisting on the conclusion of no consensus would be the way to go.
They can publish on the individual stream. The bar to cipher suite numbers is only "specification required" ffs.

2026-03-19 02:30:11 UTC

- reply

@nprofile…mv7s I totally agree but from what I've seen of what these people are doing I'm a bit jaded and not super hopeful about getting useful things out of it
The problem is also that building e.g. a Quadro card like you're describing would still incur fresh development costs and vendors just won't want to take those if the AI bubble is gone

2026-03-19 02:13:09 UTC

- reply

@nprofile…mv7s I'm afraid a good percentage will be straight up junk, i.e. big fat very-low-precision matrix multipliers… 😢 … the network gear should indeed be reusable though

2026-03-19 01:55:04 UTC

"128*400G" … these scales 😵‍💫 … all just for some slop …
https://assets.chaos.social/media_attachments/files/116/253/345/792/473/419/original/9175c2236dfecde3.jpg

2026-03-03 17:14:18 UTC

For the record, this is the IPv4 NAT performance of a Mellanox SN2010. The fact that it only has a single interrupt/queue is the major limiting factor, it's maxing out one of the 4 CPU cores while the other 3 twiddle their thumbs.
(The upload performance is probably also limited by Firefox not being able to send enough.)
IPv6, being hardware line rate switched, of course has none of these problems…
https://assets.chaos.social/media_attachments/files/116/166/350/471/055/934/original/90ff01e88dfc5bcb.png
https://assets.chaos.social/media_attachments/files/116/166/362/516/675/601/original/42060f63b3c162bd.png

2026-03-02 00:58:45 UTC

- reply

@nprofile…mv7s @nprofile…xr5j @nprofile…vnlz you can 'git bundle' the thing and hand the resulting file into the CI runner; if it doesn't need to pull dependencies (or if you can include those) you can have runners entirely without internet access.
(I demoed this at work but we deemed it excessive particularly since it makes debugging the CI rather annoying.)

2026-02-27 13:24:06 UTC

This whole business about deprecating the old ISC dhcp client would go much better if there was a fucking usable and *unobtrusive* alternative.
Like, yeah, thank you dhcpcd for fucking with half of my /proc/sys/net/ipv[46]/conf settings, I DID NOT ASK FOR ANY OF THAT.

2026-02-21 03:09:13 UTC

- reply

@nprofile…mv7s this sounds like an integration problem? Phase at (T) is "just" the fractional part of ∫ [0…T] f(t)dt and whether you can calculate that depends on what f(t) is specified as(?) I guess if your input f(t) is an arbitrary waveform, there's some math magic™ to apply… it's been a while since analysis at uni… 🤔 … I might be thoroughly confused/misunderstanding something

2026-02-05 01:43:09 UTC

- reply

@nprofile…mv7s if anything it might matter due to "software" reasons of GPU unavailability, I'm thinking RDP/VNC sessions. (Can imagine that in a lab setup even onsite…) But AVX2 should cover that, 512… meh.

2026-01-20 15:37:46 UTC

- reply

@nprofile…tpfq only my personal traffic; I was prototyping this to install it at our tiny hacker ISP.
Unfortunately the IPv6 numbers for that ISP will be abysmal because it pains me to say we haven't rolled out IPv6 to most customers since we don't want to assign static prefixes, and there's neither good FOSS tooling for randomizing & rotating prefixes, nor for actually just straight up DHCPv6-PD relay snooping. (I'm working to get the latter done in FRR.)

2026-01-20 15:17:32 UTC

Updated #IPv6 wall of shame to use @nprofile…tpfq 's ASN list, and to merge some multi-AS entities. (And counter-applied the 5% packet sampling rate.)
A little skewed by me torrenting some, uh, Linux distros 😂
Gonna leave this running for a while, let's see what falls out of it. NB: really subjective to my internet usage, not claiming this is representative in any way.
67.1% total IPv6 btw!
https://assets.chaos.social/media_attachments/files/115/928/074/222/969/909/original/4dbacf2c1c6bf29f.png

2026-01-18 05:07:16 UTC

- reply

@nprofile…0adl …und genau deswegen war es eine gute Idee Signal zu empfehlen ohne das Fass aufzumachen für wen jetzt genau wann Threema oder Telegram oder Brieftauben besser sind. Signal deckt 95% ab, und für die restlichen 5% kann man dann individuell gucken ohne die 95% ins Chaos (yes) zu reißen.

2026-01-16 04:05:35 UTC

- reply

@nprofile…mv7s not having 1500 is normal for a VPN though? I'm confused.

2026-01-16 03:47:42 UTC

- reply

@nprofile…mv7s people breaking PMTUD isn't quite T-Mobile's fault though, not that they're great or something, but on that one I really want to put the blame on the idiots actually breaking the internet with their dropping/non-handling ICMP.

2026-01-12 03:06:27 UTC

- reply

@nprofile…hf3e @nprofile…g8vl I love the UKCA conformity mark on this pile of wonky solder joints 😂

2026-01-09 23:08:06 UTC

- reply

@nprofile…9daw @nprofile…hnca ACK, a situation not improved by the lecture room situation & expectations being fuzzy :(

2026-01-09 17:23:20 UTC

- reply

@nprofile…9daw while you are mostly correct, the lecture halls have different rules again due to the VOC recording most talks. There is/was a large sign before you enter the room saying that anyone in the room may end up on a recording.
What this means for others taking photos, I don't know, and it's a bit annoying that this isn't even mentioned in the photo policy.
@nprofile…hnca

2025-12-19 05:39:17 UTC

- reply

@nprofile…syqt (I also spent 3 hours on that design hunting down a phantom 0.65V on the 1.8V permanent (ultra low idle current, <1uA) rail, until I turned off my very bright desk light and then it was gone… stupid green LED moonlighting as solar cell 😂)

2025-12-19 05:34:23 UTC

- reply

@nprofile…syqt do you have any tricks/tools/workflow for dealing with leakage between the rails? That really wasn't fun on the one 7 rail design I did, especially 3.3V permanent (off Li-Ion) vs. switched vs. externally sourced 😟

2025-12-03 23:18:12 UTC

- reply

@nprofile…syqt 'full' OSes just map all config space from all devices and leave it, especially on >=40bit virtual address space platforms. It was even a thing on i686. You never know if a PCIe switch or bridge gets involved (even hidden in a single fleck of silicon) and remapping is annoying.

2025-11-25 19:07:45 UTC

- reply

@nprofile…syqt I'm a C person, the condition of an if clause unfortunately can't declare variables… and destructors/cleanup are WIP for the next C revision… so same thing with a for loop and GCC extensions:
https://github.com/FRRouting/frr/blob/b5919324606dae54d022e2e536d55564f113a782/lib/frr_pthread.h#L232-L235

2025-11-25 18:55:16 UTC

- reply

@nprofile…syqt I have not 😮

2025-11-25 18:51:24 UTC

- reply

@nprofile…syqt we have named function arguments at home!
Named function arguments at home:
struct args {
int a, b, c, d;
};
int myfunc(struct args args);
\#define myfunc(...) myfunc((struct args){ __VA_ARGS__ })
use as:
myfunc(1, 2, .d = 4);
You're welcome! (For the new reason to scream into the abyss. I don't know where the brain bleach is, sorry.)

2025-11-22 19:39:39 UTC

- reply

@nprofile…syqt @nprofile…cx5s I'm speaking from slightly hazy memory, but I thought the file comes with newlib rather than g++… it's been a while since I built newlib but AFAIR I just built it normally, it passes a bunch of "-nostdlib" and related while building, and then you get your package including the nano spec.
Checking on my Debian install, on 32bit the gcc package only has basic CPU .specs files, and "libnewlib-arm-none-eabi" has newlib includes, libs *and* specs

2025-11-22 19:17:03 UTC

- reply

@nprofile…syqt @nprofile…cx5s I thought the specs file is all you need for that (and it has worked for me so far, though I only do embedded stuff for fun once in a while)
It seemed mostly an issue of picking the correct linker scripts and crt*.o files where necessary, and those are in the .specs…

2025-11-22 18:11:02 UTC

Ah, yes, of course the country pushing chat control in the EU would also be the one with "parenting tests", taking away newborn babies if you fail. There's no bias at all in those tests, pinky promise!
https://www.bbc.co.uk/news/articles/c1wlw2qj113o

2025-11-08 15:40:10 UTC

- reply

@nprofile…n34d I understand sometimes you need an USB device, but personally speaking I go out of my way to find TB/USB4/PCIe adapters (Apple had one for 1G). The USB stuff always has shitty bugs, especially Realtek (if I have to, I prefer ASIX for 'classic' USB).
Considering this is the 1st 10GE adapter they're making, I'd be extremely careful re. bugs (if anything doesn't work it'll likely be the adapter's fault), and performance is probably not gonna be great.

2025-10-21 08:57:56 UTC

Congrats, #systemd, you made it to the RIPE91 stage…
"every systemd-based Linux distro should reconsider using timesyncd, and switch to more robust alternatives."
https://assets.chaos.social/media_attachments/files/115/411/325/370/520/070/original/65f6d3851bfa224c.jpg

2025-10-06 10:42:45 UTC

@nprofile…em0x @nprofile…0hj7 vmtl. schon gesehen, aber Südkorea hat das mit "1x alles anzünden" etwas falsch verstanden…
https://koreajoongangdaily.joins.com/news/2025-10-01/national/socialAffairs/NIRS-fire-destroys-governments-cloud-storage-system-no-backups-available/2412936
https://www.chosun.com/english/national-en/2025/10/02/FPWGFSXMLNCFPIEGWKZF3BOQ3M/
"However, due to the system’s large-capacity, low-performance storage structure, no external backups were maintained — meaning all data has been permanently lost."

2025-06-18 12:36:06 UTC

I accidentally looked at architecture specific details on atomic ops and now I know that while cmpxchg16b is guaranteed on RAM, it can (theoretically) fail half-executed on NVDIMMs if there's a power failure…
Can I un-learn this somehow? I'm not that young anymore and need the brain space 🙃

2025-06-18 07:54:16 UTC

- reply

@nprofile…8x2j ah right the 25G ones have those tiny stubby short press fit pins for data…
Yeah sorry this sucks and I don't think my comments are useful, sorry for wasting your time 😅

2025-06-18 07:47:59 UTC

- reply

@nprofile…8x2j hmm. Right, I *think* I saw a 2-high connector separate from cage just once, might be a 'boutique' part (or I'm imagining it entirely…)
Soldering the press-fit parts is not an option? (I have nowhere near your understanding on what that does to HF behaviour…)

2025-06-18 07:44:37 UTC

- reply

@nprofile…8x2j actually… is that an integrated cage + connector? I should be clear I did this with separate cages that are only mechanical parts. But I thought the integrated cage/connectors also come in THT solder?

2025-06-18 07:42:44 UTC

- reply

@nprofile…8x2j my personal approach is to slightly under-size the holes, just going by bottom end of specified tolerance (and not pay extra for a "press-fit" option if the fab has one), and then force in the connector.
Some of the press-fit "springs" *will* break, but honestly who cares?

2025-06-15 17:22:46 UTC

- reply

@nprofile…wwe7 is this compatible in some way with code written for the same X11 feature? Does it also support keyboards? (I have some code for barcode scanners that still requires X11 because this was a missing feature in Wayland…)

2025-05-28 22:09:51 UTC

When you really, Really, REALLY, REALLLYYYYYY!!! want SSL support…
https://assets.chaos.social/media_attachments/files/114/587/737/491/293/830/original/0a9e44e4e1e1e014.png

2025-05-27 10:58:32 UTC

- reply

@nprofile…8x2j https://www.vibra-tite.com/threadlockers/medium-strength-removable/hot-lock-extreme-high-temperature-threadlocker/
(found with random googling, no prior personal experience, hopefully will never make that experience 🤣)

2025-03-31 17:19:39 UTC

"People are wrong on IETF mailing lists" is really the end boss of "people are wrong on the internet"…
Like, how many times and in how many ways am I supposed to explain how something works, both in reality as well as in established RFCs (which, in this case, easy to misread) before people run off and build things on top of plain incorrect understanding?
(I am completely, absolutely, totally certain I am right. I wrote code. I understand the code. I understand the RFC. Benefit in discourse: 0)

2025-03-30 09:14:01 UTC

- reply

@nprofile…p7ts you only need this particular fork if you don't have access to hardware perf counters. The version with perf counters is a bit faster. Also note that multithreaded things work but are serialised into a single thread (i.e. emulated task switching)

2025-03-25 16:10:44 UTC

- reply

@nprofile…p7ts stupid question from someone building software and packaging it: do you just have a different package manager than apt, or none at all?
FRRouting doesn't depend on a lot (<10 things) but it does have these deps (and a container is not viable*, and static binaries might address some bits but don't solve kernel dependencies)
* FRRouting is occasionally used to build connectivity for those very containers, containerising it is… unhelpful

2025-03-19 09:24:25 UTC

BGP over HTTP(/3, with QUIC) was just suggested on mic at IETF 🤣
(To be clear, it was effectively laughed out.)

2025-02-20 17:52:44 UTC

- reply

@nprofile…htaz don't have them in hand yet, but yeah that's one of the reasons I bought 50 of them… I need maybe 10 😅
(also the alternative would've been buying 10 for 110€…)

2025-02-20 17:39:42 UTC

anyone need CWDM4 (2km, SMF) QSFP28s? Just bought 50 of them on eBay… 150€… total 😯
…man this shit has gotten cheap.
(used of course, but still)

2025-02-10 15:02:21 UTC

I put the (public, posted here) RIPE PR photo of me on https://theyseeyourphotos.com/ and it's vaguely on point, …
… EXCEPT …
… it wants to target me for "Personalized NFT collectibles (OpenSea)"!!!!!!!1!1!1one!eleven
I have never been so offended! 😱 🤣

2025-01-30 23:39:49 UTC

big endian? thicc endian.
https://assets.chaos.social/media_attachments/files/113/919/941/849/422/981/original/2026d5f3c5e96bd8.png

2024-12-28 15:06:17 UTC

Anyone have an adapter to M.2 SATA? (from PCIe slot, SATA connector or USB) #38c3